with-line-profiler

Two independent profiling tools in one distribution: line-by-line tracing for a region you suspect, and low-overhead phase accounting for a run too long to trace. Zero dependencies on Python 3.11+, MIT licensed.

pip install with-line-profiler

# example.py
from lineprofiler import LineProfiler

def slow_function():
    total = 0
    for i in range(1_000_000):
        total += i * i
    return total

profiler = LineProfiler()
with profiler:
    slow_function()
profiler.print_stats()

====================================================================================================
File: /path/to/example.py
Function: slow_function at line 3
Total time: 710676.0 µs
====================================================================================================
Line #   Hits       Time (µs)       Per Hit (µs)    % Time     Line Content
----------------------------------------------------------------------------------------------------
6        1000000    364848.9        0.4             51.3       total += i * i
5        1000001    345825.8        0.3             48.7       for i in range(1_000_000):
7        1          0.7             0.7             0.0        return total
4        1          0.6             0.6             0.0        total = 0

(numbers vary by machine; File: prints the absolute path)

No decorators to add, no separate kernprof run, no build step. Only code under your project folder is traced — the folder is auto-detected by walking up to the nearest .git — so the output is your code, not the standard library.

Want a picture instead? profiler.to_html("profile.html") writes an annotated, heat-coloured source view as a single self-contained file.

Which tool

They share nothing but the distribution: accounting never imports LineProfiler. If you arrived here for a training run, you want the accounting layer — it is the one built to stay enabled for twelve hours.

from lineprofiler.accounting import start, stop

start(role="actor")     # or Profiler(...) / with profiler.phase(...)
...
stop()

lineprofiler report profile/                        # the run, as a table
lineprofiler report profile/ --format html -o r.html # ...or as a page you can share

Documentation

• The line profiler — the with block, start_profiling(), and what it does not do
• The accounting layer — phases, counters, and instrumenting without threading an argument
• Accounting recipes — reading the report, I/O and GPU bottlenecks, overhead budgets, exporting to W&B
• Multiple processes and nodes — Slurm, forking, preemption, heavy backends
• HTML reports — the icicle chart, the annotated source view, and the embedded data block
• Configuration — environment variables, [tool.lineprofiler], optional dependencies
• Comparison with other profilers — line_profiler, py-spy, Scalene, VizTracer, and when to use those instead

Python support

3.10 and newer. On 3.12+ the line profiler uses sys.monitoring, so it can run alongside coverage.py, pdb and other tracing tools; below that it falls back to sys.settrace, which is a single global hook and cannot. tomli is required only on 3.10, where tomllib is not yet in the standard library.

Licence

MIT

The claude.md is partially created from https://github.com/multica-ai/andrej-karpathy-skills/blob/main/CLAUDE.md